Lecture 14: Distributed Learning, Security, and Privacy

transcript

1Serena Yeung BIODS 220: AI in Healthcare Lecture 14 -

Announcements● TA office hours will be project advising sessions during this week and the

week of the 29th○ Sign up on spreadsheet (see Ed announcement)○ Attendance is worth 5% of project grade

● No class or OH next week due to university holiday● Wed 11/17 (over zoom): Guest lecture: Prof Julia Salzman (molecular

discovery, single cell sequencing)○ Extra credit opportunity!

● Mon 11/29: Course conclusion

Agenda- Distributed Learning and Federated Learning- Privacy and Differential Privacy

Distributed Learning

Figure credit: Alsheikh et al. Mobile big data analytics using deep learning and apache spark, 2016.

- Sharing the computational load of training a model among multiple worker nodes

Data and task of computing gradient updates is distributed among nodes

Can have data parallelism or model parallelism

Federated Learning- Related to distributed computing, but with an important property for many medical

settings: data is decentralized and never leaves local silos. Central server controls training across decentralized sources.

Figure credit: https://blogs.nvidia.com/wp-content/uploads/2019/10/federated_learning_animation_still_white.png

Federated Learning

Figure credit: https://blog.ml.cmu.edu/2019/11/12/federated-learning-challenges-methods-and-future-directions/

- Example: learning a next-word prediction model from many individual cell phones

Federated Learning

Current copy of global model is shipped to local devices that are ready to contribute to training

Federated Learning

Local gradient updates are computed on local data after 1 or several iterations of gradient descent

Federated Learning

Local gradient updates are shipped back to the central server

Federated Learning

Local gradient updates are combined and used to update the global model

Federated Learning

Updated global model is shipped to local devices for future rounds of federated learning training

Federated Learning

- Example: learning a personalized healthcare model from data across different healthcare organizations

From earlier: BRATS brain tumor segmentation dataset- Segmentation of tumors in brain MR image slices- BRATS 2015 dataset: 220 high-grade brain tumor and 54 low-grade brain tumor MRIs- U-Net architecture, Dice loss function

Dong et al. Automatic Brain Tumor Detection and Segmentation Using U-Net Based Fully Convolutional Networks. MIUA, 2017.

Li et al. 2019- NVIDIA Clara’s Federated Learning system for medical imaging data

- Used federated learning to train segmentation model on BRATS

- Achieved comparable performance to non-federated learning, training somewhat slower but data “silos” preserved

Li et al. Privacy-preserving Federated Brain Tumour Segmentation, 2019.

Li et al. 2019- NVIDIA Clara’s Federated Learning system for medical imaging data

- Used federated learning to train segmentation model on BRATS

- Achieved comparable performance to non-federated learning, training somewhat slower but data “silos” preserved

Li et al. Privacy-preserving Federated Brain Tumour Segmentation, 2019.

Also differentially private version… will talk about this in a moment

Privacy: HIPAA

Figure credit: https://www.jet-software.com/en/data-masking-hipaa/

Health Insurance Portability and Accountability Act (HIPAA), 1996: created “Privacy Rule” for how healthcare entities must protect the privacy of patients’ medical information

18 HIPAA identifiers (Protected Health Information):

Risks of data re-identification

Figure credit: Sweeney et al. Matching Known Patients to Health Records in Washington State Data, 2011.

Data triangulation: a person may be de-identified as to one data set, but the knowledge that they are a member of another available data set may allow them to be reidentified

Matching Known Patients to Health Records in Washington State Data

Sweeney. Matching Known Patients to Health Records in Washington State Data, 2011.

News stories (e.g., those containing the word “hospitalized”) contain identifying information that could be used to identify medical records in the state medical record database, for 43% of studied cases

Distribution of values for fields harvested from news stories

Identifying Participants in the Personal Genome Project by Name

Sweeney. Identifying Participants in the Personal Genome Project by Name, 2013.

Linked demographics information in the Personal Genome Project (PGP) to public records such as voter lists, to correctly identify 84 to 97% of profiles for which guessed names were provided to PGP staff

K-anonymityA data release provides k-anonymity protection if the information for each person contained in the release cannot be distinguished from at least k-1 individuals whose information also appears in the release.

Sweeney. K-anonymity: a model for protecting privacy. 2002.

K-anonymity

Sweeney. K-anonymity: a model for protecting privacy. 2002.

2 k-anonymity tables (where

k = 2)

Re-identification from ML models

Fredrickson et al. Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures, 2015.

- White-box (as opposed to black-box) setting: have access to model parameters, e.g. local model downloaded on device to run inference

- Model inversion attack: can use gradient descent if model parameters are available, to infer sensitive features

Differential privacyKey idea: output for a dataset, vs. the dataset with a difference for a single entry (e.g., one individual), is “hardly different”. Mathematical guarantees on this idea.

Abadi et al. Deep Learning with Differential Privacy, 2016.

Differential privacyKey idea: output for a dataset, vs. the dataset with a difference for a single entry (e.g., one individual), is “hardly different”. Mathematical guarantees on this idea.

Differential privacySimple intuition behind how we can achieve differential privacy: adding noise!

Figure credit: https://github.com/frankmcsherry/blog/blob/master/posts/2016-02-03.md

Example of reporting a value with Laplacian noise added

Training differentially private deep learning models

Compute gradient as usual

Clip the gradient

Add noise for differential privacy

Compute overall privacy cost

Privacy Aggregation of Teacher Ensembles (PATE)

Papernot et al. Semi-supervised Knowledge Transfer for Deep Learning from Private Data, 2017.

PATE Teacher model

Approach to combine data from multiple disjoint sensitive populations, with privacy guarantees

Figure credit: http://www.cleverhans.io/privacy/2018/04/29/privacy-and-machine-learning.html

PATE Teacher model

Approach to combine data from multiple disjoint sensitive populations, with privacy guarantees Train separate classifiers from

disjoint data sets -- no privacy guarantees yet

PATE Teacher model

Approach to combine data from multiple disjoint sensitive populations, with privacy guarantees To get a privacy-preserving

prediction, first obtain predictions from all distinct classifiers

PATE Teacher model

Approach to combine data from multiple disjoint sensitive populations, with privacy guarantees Then add noise to the vote

histogram (giving differential privacy guarantees), and take the class with the most votes as the final output

PATE Student model uses public data to train a model replicating noisy aggregated teacher outputs

This teacher model alone can still be compromised if too many queries are performed (privacy cost builds up with each query, so privacy guarantees become meaningless with too many queries), or if model parameters are made accessible (and attackable) e.g. distributed in local application

Generator network: try to fool the discriminator by generating real-looking imagesDiscriminator network: try to distinguish between real and fake images

zRandom noise

Generator Network

Discriminator Network

Fake Images(from generator)

Real Images(from training set)

Real or Fake

After training, use generator network to generate new images

Ian Goodfellow et al., “Generative Adversarial Nets”, NIPS 2014

Fake and real images copyright Emily Denton et al. 2015. Reproduced with permission.

Remember GANs: Two-player game

Generator network: try to fool the discriminator by generating real-looking imagesDiscriminator network: try to distinguish between real and fake images

zRandom noise

Generator Network

Discriminator Network

Fake Images(from generator)

Real Images(from training set)

Real or Fake

After training, use generator network to generate new images

Ian Goodfellow et al., “Generative Adversarial Nets”, NIPS 2014

Fake and real images copyright Emily Denton et al. 2015. Reproduced with permission.

Remember GANs: Two-player game

Can train GANs using differentially private SGD (DP-SGD)! Afterwards, using the generator to generate synthetic data does not incur additional privacy cost

Xie et al. Differentially Private Generative Adversarial Network, 2018.

Implementation of DP-SGD

Utilities for calculating epsilon

Can work with differential privacy within deep learning frameworks

https://blog.tensorflow.org/2019/03/introducing-tensorflow-privacy-learning.htmlhttp://www.cleverhans.io/privacy/2019/03/26/machine-learning-with-differential-privacy-in-tensorflow.html

Today we covered:- Distributed Learning and Federated Learning- Privacy and Differential Privacy

Next time: Guest lecture from Prof. Julia Salzman over zoom, discussing molecular discovery and single cell sequencing

Lecture 14: Distributed Learning, Security, and Privacy

Documents